A fast computational algorithm for and performance of the Kumaresan-Prony method of spectrum analysis

Kumaresan and Tufts at the University of Rhode Island have developed a modification of the Prony technique that yields improved linear prediction frequency estimates for two equiamplitude sinusoids in white noise. The improved performance is manifested by decreased variance and bias in the estimates of the two closely-spaced sinusoid frequencies, particularly at low signal-to-noise ratios (SNR) where other linear prediction techniques do not seem to do as well. The Kumaresan-Prony technique is based on fitting a high-order linear prediction model to the available data using a pseudoinverse solution. This paper indicates how a fast computational algorithm may be obtained to solve the pseudoinverse matrix. Also, an examination of the spectral estimation performance against a different signal scenario is made.